UpdatingEMS: An online updating framework for deep reinforcement learning-based energy management of fuel cell hybrid electric bus with integrated transfer learning

Deep reinforcement learning (DRL) holds great promise in enhancing the effectiveness of energy management strategies (EMSs) for hybrid electric vehicles (HEVs). However, online updating of the DRL-based EMSs remains a challenge, making it difficult to ensure their long-term optimization performance. Given that, this study proposes an online updating EMS to improve the long-term energy efficiency of the DRL-based EMS for a fuel cell hybrid electric bus, by exploiting the correlation mechanism between real-time traffic information and efficient hydrogen utilization. Specifically, future optimal safety speed is planned by adopting dynamic programming addressing coupled spatiotemporal constraints in traffic information. Furthermore, a knowledge-sharing mechanism is developed by leveraging transfer learning (TL) to reuse historical EMS for the planned future speed, enabling the continuous updating of the soft actor-critic based EMS. Finally, the updated EMS is deployed into the onboard controller to verify the real-time control effect via the processor-in-the-loop experiment. Results demonstrate that the proposed EMS enhances updating efficiency by 30.08 % compared to the non-TL-integrated EMS and reduces hydrogen consumption by 6.11 % compared to the static EMS. Moreover, the updated EMS can be deployed in real time in the onboard controller. This study is a proof-of-concept demonstrating the EMS updating under idealized assumptions, making a contribution to intelligent transportation technologies on energy conservation from the aspect of energy management for HEVs